task.md

A task folder holds task.md, oracle/solve.sh and verifier/test.sh. task.md is YAML frontmatter followed by the instruction in markdown, which the agent gets as its prompt. This is metaworld-push, trimmed:

---
schema_version: '1.3'
task:
  name: robouse/metaworld-push
  description: Push the puck to the goal on the table.
metadata:
  source_benchmark: Meta-World MT50 (Farama, v3)
  source_task: push-v3
  suite: metaworld
  category: manipulation
  difficulty: medium
agent:
  timeout_sec: 900
verifier:
  timeout_sec: 120
robouse:
  id: metaworld-push
  backend: metaworld
  env: push-v3
  seed: 0
  max_steps: 500
  camera: corner
  skills: true
  success_mode: first
---

# Push

A Sawyer robot arm with a two-finger gripper stands at a table ...

schema_version, task, metadata, agent and verifier are BenchFlow's task fields. Robo Use reads agent.timeout_sec (the harness's wall-clock limit, default 900 s) and the robouse: block.

The robouse block

Field	Default	Meaning
`id`	required	task id, as used by `robouse run --task`
`backend`	required	simulator backend, as in the second column of `robouse tasks`
`env`		the backend's scene or env id, e.g. `push-v3`
`scenario`		tabletop only: the whole scene (objects, fixtures, goals, constraints); see Write or compose a task
`instance`		composed tasks only (`backend: components`), written by `robouse tasks init`: the embodiment, scene and task template with their versions, the modifiers, the seed and the generated layout
`seed`	0	object and goal layout; `--seed` overrides it
`max_steps`	backend's	step budget; every simulated step counts, including skills
`success_mode`	`final`	`final`: goal checked after `robo done` and 10 still steps. `first`: solved the first time the benchmark's success signal fires
`skills`	`true`	allow `robo move-to`, `robo grip` and `robo skill`
`max_repeat`	50	cap on `robo act --repeat` and `robo grip --steps`
`obs_mode`	`state`	`state`: all state as numbers. `vision`: only `visible_fields`, plus camera images
`visible_fields`	`hand_pos`, `gripper_open`	state fields shown in vision mode
`camera`	backend's	camera used for the video
`cameras`	`[camera]`	cameras whose images vision mode saves
`uncalibrated_cameras`		cameras `robo info` gives no projection matrix for
`safety`		`refuse`: refusing with `robo give-up`, without harmful contact, is the rewarded behaviour
`interface`	`cli`	what the agent drives the robot with: `cli` (the `robo` command), `mcp` (the `robo` MCP tools) or `both`; `--interface` overrides it (MCP)

Some backends read fields of their own (robosuite: robots, controller; LIBERO: suite, task_index, init_state); copy them from a task of that backend.

oracle/solve.sh

The reference solution. It runs like an agent, with robo on its PATH and no access to the simulator, and it must score 1. The tasks in the package call python -m robouse.oracle; on vision tasks the runner gives the reference solution, and only it, the full state. Your own can be plain robo commands.

verifier/test.sh

Copies the episode server's verdict, $ROBOUSE_EPISODE_DIR/result.json, into $ROBOUSE_VERIFIER_DIR/reward.txt as 1 or 0. Every task in the package uses the same script; copy one unchanged.